feat(llm): add Atlas Cloud router

.env.public

@@ -19,6 +19,10 @@ LITELLM_BASE_URL=
 TOKENROUTER_API_KEY=
 TOKENROUTER_BASE_URL=https://api.tokenrouter.com/v1
 
+# ── Atlas Cloud (OpenAI-compatible; atlascloud/* models) ──────────────────────────────
+ATLASCLOUD_API_KEY=
+ATLASCLOUD_API_BASE=https://api.atlascloud.ai/v1
+
 # ── Benchmark code execution
 SCENARIO_DIR=
 LEADERBOARD_DIR=

INSTRUCTIONS.md

@@ -126,6 +126,13 @@ See [MCP Servers](#mcp-servers) for available tools and [docs/mcp-servers.md](do
 | `TOKENROUTER_API_KEY`  | _(tokenrouter/* models)_ | TokenRouter API key                                                 |
 | `TOKENROUTER_BASE_URL` | _(tokenrouter/* models)_ | TokenRouter base URL, e.g. `https://api.tokenrouter.com/v1`         |
 
+**Atlas Cloud** — OpenAI-compatible gateway for the direct LLM baseline, selected by the `atlascloud/` prefix
+
+| Variable              | Default                         | Description                                      |
+| --------------------- | ------------------------------- | ------------------------------------------------ |
+| `ATLASCLOUD_API_KEY`  | _(atlascloud/* models)_          | Atlas Cloud API key                              |
+| `ATLASCLOUD_API_BASE` | `https://api.atlascloud.ai/v1` | Atlas Cloud API base URL (optional)              |
+
 **OpenCode direct providers** — `opencode-agent` with direct provider routes
 
 | Variable            | Default                    | Description                                      |
@@ -316,6 +323,11 @@ uv run opencode-agent \
 # Direct model-only baseline, no MCP tools
 uv run direct-llm-agent --model-id litellm_proxy/Azure/gpt-5-mini-2025-08-07 \
   'Return only JSON: {"test": 1}'
+
+# Direct model-only baseline through Atlas Cloud
+ATLASCLOUD_API_KEY=... uv run direct-llm-agent \
+  --model-id atlascloud/openai/gpt-5.4 \
+  'Return only JSON: {"test": 1}'
 ```
 
 ### OpenCode scenario-suite workspace mode

src/llm/openai_compat.py

@@ -1,7 +1,8 @@
 """OpenAI-compatible LLM backend (no litellm dependency).
 
 For gateways that expose the standard OpenAI Chat Completions API — such as
-`TokenRouter <https://www.tokenrouter.com>`_ — we talk to them with the
+`TokenRouter <https://www.tokenrouter.com>`_ and
+`Atlas Cloud <https://www.atlascloud.ai>`_ — we talk to them with the
 ``openai`` SDK directly instead of routing through litellm.  litellm only
 earns its keep for providers that are *not* OpenAI-shaped (e.g. watsonx).
 
@@ -15,7 +16,12 @@
 from __future__ import annotations
 
 from .base import LLMBackend, LLMResult
-from .routers import is_openai_compat, resolve_model, resolve_router_creds
+from .routers import (
+    ATLASCLOUD_PREFIX,
+    is_openai_compat,
+    resolve_model,
+    resolve_openai_compat_creds,
+)
 
 __all__ = ["OpenAICompatBackend", "is_openai_compat"]
 
@@ -39,8 +45,19 @@ def generate(self, prompt: str, temperature: float = 0.0) -> str:
     def generate_with_usage(self, prompt: str, temperature: float = 0.0) -> LLMResult:
         from openai import OpenAI
 
-        creds = resolve_router_creds(self._model_id)  # strict: clear error if unset
-        client = OpenAI(base_url=creds.base_url, api_key=creds.api_key)
+        creds = resolve_openai_compat_creds(self._model_id)
+        if creds is None:
+            raise ValueError(
+                f"missing OpenAI-compatible credentials for {self._model_id!r}"
+            )
+        if creds.prefix == ATLASCLOUD_PREFIX:
+            client = OpenAI(
+                base_url=creds.base_url,
+                api_key=creds.api_key,
+                max_retries=0,
+            )
+        else:
+            client = OpenAI(base_url=creds.base_url, api_key=creds.api_key)
         response = client.chat.completions.create(
             model=self._model_name,
             messages=[{"role": "user", "content": prompt}],

src/llm/routers.py

@@ -12,6 +12,7 @@
 
     litellm_proxy/<model>   LiteLLM proxy  (LITELLM_BASE_URL / LITELLM_API_KEY)
     tokenrouter/<model>     TokenRouter    (TOKENROUTER_BASE_URL / TOKENROUTER_API_KEY)
+    atlascloud/<model>      Atlas Cloud    (ATLASCLOUD_API_KEY)
 """
 
 from __future__ import annotations
@@ -21,6 +22,8 @@
 
 LITELLM_PREFIX = "litellm_proxy/"
 TOKENROUTER_PREFIX = "tokenrouter/"
+ATLASCLOUD_PREFIX = "atlascloud/"
+ATLASCLOUD_DEFAULT_BASE_URL = "https://api.atlascloud.ai/v1"
 
 
 class RouterCreds(NamedTuple):
@@ -39,12 +42,12 @@ class RouterCreds(NamedTuple):
 
 # Prefixes whose endpoints speak the OpenAI Chat Completions API and can be
 # driven by the native ``openai`` SDK (llm.OpenAICompatBackend).
-OPENAI_COMPAT_PREFIXES: tuple[str, ...] = (TOKENROUTER_PREFIX,)
+OPENAI_COMPAT_PREFIXES: tuple[str, ...] = (TOKENROUTER_PREFIX, ATLASCLOUD_PREFIX)
 
 
 def router_prefix(model_id: str) -> str | None:
     """Return the proxy-router prefix matching *model_id*, else ``None``."""
-    for prefix in PROXY_ROUTERS:
+    for prefix in (*PROXY_ROUTERS, ATLASCLOUD_PREFIX):
         if model_id.startswith(prefix):
             return prefix
     return None
@@ -65,6 +68,23 @@ def is_openai_compat(model_id: str) -> bool:
     return model_id.startswith(OPENAI_COMPAT_PREFIXES)
 
 
+def resolve_openai_compat_creds(model_id: str) -> RouterCreds | None:
+    """Resolve credentials for native OpenAI-compatible backends."""
+    if model_id.startswith(ATLASCLOUD_PREFIX):
+        api_key = os.environ.get("ATLASCLOUD_API_KEY")
+        if not api_key:
+            raise ValueError(
+                "ATLASCLOUD_API_KEY must be set when using the 'atlascloud/' model prefix"
+            )
+        return RouterCreds(
+            prefix=ATLASCLOUD_PREFIX,
+            base_url=os.environ.get("ATLASCLOUD_API_BASE")
+            or ATLASCLOUD_DEFAULT_BASE_URL,
+            api_key=api_key,
+        )
+    return resolve_router_creds(model_id)
+
+
 def resolve_router_creds(model_id: str, *, strict: bool = True) -> RouterCreds | None:
     """Resolve endpoint + key for *model_id*, or ``None`` if not proxied.
 
@@ -74,7 +94,7 @@ def resolve_router_creds(model_id: str, *, strict: bool = True) -> RouterCreds |
             ``None`` so the caller can fall back to its own defaults.
     """
     prefix = router_prefix(model_id)
-    if prefix is None:
+    if prefix is None or prefix not in PROXY_ROUTERS:
         return None
     base_env, key_env = PROXY_ROUTERS[prefix]
     base_url = os.environ.get(base_env)

src/llm/tests/test_backends.py

@@ -23,9 +23,10 @@ def create(**kwargs):
         )
 
     class OpenAI:
-        def __init__(self, base_url=None, api_key=None):
+        def __init__(self, base_url=None, api_key=None, **kwargs):
             captured["base_url"] = base_url
             captured["api_key"] = api_key
+            captured.update(kwargs)
             self.chat = types.SimpleNamespace(
                 completions=types.SimpleNamespace(create=create)
             )
@@ -37,12 +38,14 @@ def __init__(self, base_url=None, api_key=None):
 
 def test_is_openai_compat():
     assert is_openai_compat("tokenrouter/MiniMax-M3")
+    assert is_openai_compat("atlascloud/openai/gpt-5.4")
     assert not is_openai_compat("litellm_proxy/aws/claude-opus-4-6")
     assert not is_openai_compat("watsonx/meta-llama/llama-3-3-70b-instruct")
 
 
 def test_make_backend_dispatch():
     assert isinstance(make_backend("tokenrouter/MiniMax-M3"), OpenAICompatBackend)
+    assert isinstance(make_backend("atlascloud/openai/gpt-5.4"), OpenAICompatBackend)
     assert isinstance(make_backend("litellm_proxy/aws/claude-opus-4-6"), LiteLLMBackend)
     assert isinstance(make_backend("watsonx/meta-llama/x"), LiteLLMBackend)
 
@@ -63,10 +66,28 @@ def test_tokenrouter_strips_prefix_and_routes(monkeypatch):
     assert captured["model"] == "MiniMax-M3"  # bare name, prefix stripped
     assert captured["base_url"] == "https://api.tokenrouter.com/v1"
     assert captured["api_key"] == "tr-key"
+    assert "max_retries" not in captured
     assert result.text == "hi"
     assert (result.input_tokens, result.output_tokens) == (3, 2)
 
 
+def test_atlascloud_uses_default_endpoint_without_sdk_retries(monkeypatch):
+    captured: dict = {}
+    _install_fake_openai(monkeypatch, captured)
+    monkeypatch.setenv("ATLASCLOUD_API_KEY", "atlas-key")  # pragma: allowlist secret
+    monkeypatch.delenv("ATLASCLOUD_API_BASE", raising=False)
+
+    result = make_backend(
+        "atlascloud/dots-studio/dots-3-note-prev-free"
+    ).generate_with_usage("hello")
+
+    assert captured["model"] == "dots-studio/dots-3-note-prev-free"
+    assert captured["base_url"] == "https://api.atlascloud.ai/v1"
+    assert captured["api_key"] == "atlas-key"  # pragma: allowlist secret
+    assert captured["max_retries"] == 0
+    assert result.text == "hi"
+
+
 def test_model_id_property_keeps_full_string():
     assert (
         OpenAICompatBackend("tokenrouter/MiniMax-M3").model_id

src/llm/tests/test_routers.py

@@ -5,10 +5,13 @@
 import pytest
 
 from llm.routers import (
+    ATLASCLOUD_DEFAULT_BASE_URL,
+    ATLASCLOUD_PREFIX,
     LITELLM_PREFIX,
     TOKENROUTER_PREFIX,
     is_openai_compat,
     resolve_model,
+    resolve_openai_compat_creds,
     resolve_router_creds,
     router_prefix,
 )
@@ -17,13 +20,15 @@
 def test_prefix_constants():
     assert LITELLM_PREFIX == "litellm_proxy/"
     assert TOKENROUTER_PREFIX == "tokenrouter/"
+    assert ATLASCLOUD_PREFIX == "atlascloud/"
 
 
 @pytest.mark.parametrize(
     "model_id,expected",
     [
         ("litellm_proxy/aws/claude-opus-4-6", "aws/claude-opus-4-6"),
         ("tokenrouter/MiniMax-M3", "MiniMax-M3"),
+        ("atlascloud/openai/gpt-5.4", "openai/gpt-5.4"),
         ("anthropic/claude-sonnet-4-6", "anthropic/claude-sonnet-4-6"),
         ("gpt-4o", "gpt-4o"),
         ("", ""),
@@ -38,6 +43,7 @@ def test_resolve_model(model_id, expected):
     [
         ("litellm_proxy/aws/claude-opus-4-6", "litellm_proxy/"),
         ("tokenrouter/MiniMax-M3", "tokenrouter/"),
+        ("atlascloud/openai/gpt-5.4", "atlascloud/"),
         ("anthropic/claude-sonnet-4-6", None),
     ],
 )
@@ -47,6 +53,7 @@ def test_router_prefix(model_id, expected_prefix):
 
 def test_is_openai_compat():
     assert is_openai_compat("tokenrouter/MiniMax-M3")
+    assert is_openai_compat("atlascloud/openai/gpt-5.4")
     assert not is_openai_compat("litellm_proxy/aws/claude-opus-4-6")
     assert not is_openai_compat("watsonx/meta-llama/x")
 
@@ -62,6 +69,33 @@ def test_resolve_router_creds_tokenrouter(monkeypatch):
 
 def test_resolve_router_creds_native_passthrough():
     assert resolve_router_creds("anthropic/claude-sonnet-4-6") is None
+    assert resolve_router_creds("atlascloud/openai/gpt-5.4") is None
+
+
+def test_resolve_openai_compat_creds_atlascloud(monkeypatch):
+    monkeypatch.setenv("ATLASCLOUD_API_KEY", "atlas-key")  # pragma: allowlist secret
+    monkeypatch.delenv("ATLASCLOUD_API_BASE", raising=False)
+
+    creds = resolve_openai_compat_creds("atlascloud/openai/gpt-5.4")
+
+    assert creds.prefix == ATLASCLOUD_PREFIX
+    assert creds.base_url == ATLASCLOUD_DEFAULT_BASE_URL
+    assert creds.api_key == "atlas-key"  # pragma: allowlist secret
+
+
+def test_resolve_openai_compat_creds_atlascloud_custom_base(monkeypatch):
+    monkeypatch.setenv("ATLASCLOUD_API_KEY", "atlas-key")  # pragma: allowlist secret
+    monkeypatch.setenv("ATLASCLOUD_API_BASE", "https://atlas.example/v1")
+
+    creds = resolve_openai_compat_creds("atlascloud/openai/gpt-5.4")
+
+    assert creds.base_url == "https://atlas.example/v1"
+
+
+def test_resolve_openai_compat_creds_atlascloud_requires_key(monkeypatch):
+    monkeypatch.delenv("ATLASCLOUD_API_KEY", raising=False)
+    with pytest.raises(ValueError, match="ATLASCLOUD_API_KEY"):
+        resolve_openai_compat_creds("atlascloud/openai/gpt-5.4")
 
 
 def test_resolve_router_creds_strict_raises(monkeypatch):